Pete Florence

Google DeepMind to Physical Intelligence Co-founder

Pete Florence

Home > People > Pete Florence

Profile

FieldDetails
Current PositionPhysical Intelligence Co-founder
PreviousGoogle DeepMind Research Scientist
PhDMIT
AdvisorRuss Tedrake

Key Contributions

  • Dense Object Nets: Learning dense visual descriptors for objects
  • Implicit Representations for Robotics: Applying NeRF and similar methods to robotics
  • Physical Intelligence Founding: Participated in pi0 development
  • Google Robotics Research: Visual representation learning research

Research Timeline

MIT PhD (2014-2019)

Advised by Russ Tedrake

YearWorkImpact
2018Dense Object NetsDense visual descriptors
2019PhD Graduation

Google Brain / DeepMind (2019-2024)

Visual Representation + Robot Learning

YearWorkImpact
2019Joined GoogleGoogle Brain Robotics
2020Implicit RepresentationsNeRF + Robotics
2021Transporter Networks relatedObject rearrangement
2022Continued visual representation research

Physical Intelligence (2024-present)

YearWorkImpact
2024Co-founded Physical Intelligence
2024pi0 Development participation

Major Publications

Dense Descriptors

  • Dense Object Nets (CoRL 2018) - Most influential paper
  • Self-Supervised Correspondence (2019)

Implicit Representations

  • NeRF for Manipulation (2021)
  • Implicit Behavioral Cloning (2021)

Language + Vision + Action

  • CLIPort related research
  • Language-conditioned manipulation

Key Ideas

Dense Object Nets (2018)

Core: Learning consistent descriptors for every pixel on objects

Features:
- Invariant to viewpoint changes
- Correspondence between object instances
- Self-supervised learning

Applications:
- Robot grasping
- Object rearrangement
- Manipulation tasks

Impact:

  • Core research in robot visual representation learning
  • Influenced Transporter Networks and others
  • Dense correspondence application to robotics

Implicit Representations for Robotics

Core: Utilizing implicit representations like NeRF for robot manipulation

Advantages:
- Continuous 3D representation
- Novel view synthesis
- Physical interaction prediction

Philosophy & Direction

Research Philosophy

“Good visual representation is the key to robot learning. What you see determines what you can do.”

Research Direction Evolution

  1. 2014-2019: Dense visual descriptors, self-supervised learning
  2. 2019-2022: Implicit representations, NeRF + robotics
  3. 2022-2024: Language-conditioned manipulation
  4. 2024-present: Foundation models, Physical Intelligence

MIT to Google to Physical Intelligence

MIT Experience

  • Advised by Russ Tedrake (robot control expert)
  • Experience with Drake simulator
  • Theory + practical robotics combination

Google Experience

  • Large-scale research infrastructure
  • Collaboration with RT series colleagues
  • Foundation model experience

Physical Intelligence Founding

  • From academic research to actual productization
  • Founded together with colleagues

See Also